The Null Steady-State Distribution of the CUSUM Statistic

We develop an empirical approximation to the null steady-state distribution of the cumulative sum (CUSUM) statistic, defined as the distribution of values obtained by running a CUSUM with no upper boundary under the null state for an indefinite period of time. The derivation is part theoretical and part empirical and the approximation is valid for CUSUMs applied to Normal data with known variance (although the theoretical result is true in general for exponential family data, which we show in the Appendix). The result leads to an easily-applied formula for steady-state p-values corresponding to CUSUM values, where the steady-state p-value is obtained as the tail area of the null steady-state distribution and represents the expected proportion of time, under repeated application of the CUSUM to null data, that the CUSUM statistic is greater than some particular value. When designing individual charts with fixed boundaries, this measure could be used alongside the average run-length (ARL) value, which we show by way of examples may be approximately related to the steady-state p-value. For multiple CUSUM schemes, use of a p-value enables application of a signalling procedure that adopts a false discovery rate (FDR) approach to multiplicity control. Under this signalling procedure, boundaries on each individual chart change at every observation according to the ordering of CUSUM values across the group of charts. We demonstrate practical application of the steady-state p-value to a single chart where the example data are number of earthquakes per year registering > 7 on the Richter scale measured from 1900 to 1998. Simulation results relating to the empirical approximation of the null steady-state distribution are summarised, and those relating to the statistical properties of the proposed signalling procedure for multiple CUSUM schemes are presented.